Bug report

If I deploy cilium-agent on a host that has more than 1 NIC, it is possible that the NIC used for cluster communication is not the same as the NIC used as gateway. However, Cilium will always autodetect the MTU by reading the mtu of the NIC used as gateway. As a consequence, if the MTUs are different, the detected MTU will be wrong.

This could be solved if the autodetection would take the assigned nodeIP for that cilium-agent, find the interface that holds that IP and then fetch the MTU from that interface.

General Information

• Cilium version: master
• Kernel version: Not relevant
• Orchestration system version in use: Kubernetes
• Link to relevant artifacts: https://github.com/cilium/cilium/blob/master/pkg/mtu/detect_linux.go#L49-L73

How to reproduce the issue

1. Having a server with 2 NICs, use NIC1 as gateway with mtu=1500. Use NIC2 for cluster communication with mtu=9000
2. Deploy Cilium and even if cilium-agent pods take the IP of NIC2, the autodetected mtu is 1500